Causal Inference under Dynamic Selection and Latent Heterogeneity
summary
The gist
The goal is to create a thorough overview that captures the methodological innovation, theoretical rigor, and empirical application described in the source material.
In short
This research develops a method to estimate dynamic treatment effects when treatments are adopted at different times, depending on both unobserved time-invariant factors and changing outcomes. The authors propose a novel identification strategy using latent factors and pretreatment histories, leading to a kernel-based doubly robust estimator. This technique provides consistent and unbiased estimates for the average treatment effect on the treated.
Key concepts
- Dynamic Treatment Effects (DTE)
- This measures how a treatment received at one time affects future outcomes. Because treatments are adopted staggered over time, we need a method that accounts for this changing timing to accurately estimate the true impact of receiving treatment.
- Time-Invariant Latent Factors
- These are unobserved characteristics of individuals that do not change over time. The model uses these factors to help identify individuals who are similar in their underlying traits, even when their observed pretreatment data looks different.
- Kernel-Based Doubly Robust Estimator
- This is the statistical tool used to calculate the treatment effect. It combines two flexible estimation methods: one based on covariates and another based on a sample distance metric (kernel), making it robust against model misspecification.
Terminology used across episodes
This episode discusses
- Causal Inference under Dynamic Selection and Latent Heterogeneity · Paper Radio
- Doubly Robust Inference in Causal Latent Factor Models
- Large-Sample Properties of the Synthetic Control Method under Selection on Unobservables
- Identification of Average Treatment Effects in Nonparametric Panel Models
- Back to Feedback: Dynamics and Heterogeneity in Panel Data
- Time-Varying Heterogeneous Treatment Effects in Event Studies
- Difference in Differences with Time-Varying Covariates
- Linear Estimation of Structural and Causal Effects for Nonseparable Panel Data
- Automatic Debiased Machine Learning for Dynamic Treatment Effects and General Nested Functionals
- Proxy Controls and Panel Data
- Inferring Treatment Effects in Large Panels by Uncovering Latent Similarities
- Causal Inference in Possibly Nonlinear Factor Models · Paper Radio
- Optimal Estimation of Large-Dimensional Nonlinear Factor Models
- Selection and parallel trends
- Estimating Dyadic Treatment Effects with Unknown Confounders
- Double/Debiased Machine Learning for Dynamic Treatment Effects via g-Estimation
- Heterogeneous Treatment Effects via Dynamic Panel Data Methods · Paper Radio
- Cross-Fitting and Fast Remainder Rates for Semiparametric Estimation
- Distributionally Robust Receive Combining
- Flexible Imputation of Incomplete Network Data
The paper
Causal Inference under Dynamic Selection and Latent Heterogeneity · Read on arXiv
Weisheng Zhang
University College London
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: We're looking at the paper: "Causal Inference under Dynamic Selection and Latent Heterogeneity".
Warren: Here's the gist of the paper. The goal is to create a thorough overview that captures the methodological innovation, theoretical rigor, and empirical application described in the source material.
Ruth: Here is the combined, detailed summary: Comprehensive Research Summary: Causal Inference under Dynamic Selection and Latent Heterogeneity This research paper presents a sophisticated framework for estimating dynamic treatment effects (DTE) in panel data settings characterized by staggered adoption of treatments, where treatment timing depends jointly on unobserved time-invariant heterogeneity and time-varying pretreatment covariates, including lagged outcomes.
Warren: The core challenge addressed is the nonparametric identification of the dynamic Average Treatment Effect on the Treated (ATT) under these complex selection mechanisms.
Ruth: Methodological Innovation and Identification Strategy: The central methodological contribution is a novel identification strategy that leverages time-invariant latent factors—which are unobserved—and time-varying pretreatment covariates.
Paper summary: Ruth: So we’re looking at this paper today, "Causal Inference under Dynamic Selection and Latent Heterogeneity". It tackles those tricky situations where you can't just compare groups because the timing of getting a treatment depends on things that are unobserved and change over time.
Warren: Exactly. The main thing they are doing is studying dynamic treatment effects when the selection process is complicated by both those static, unobserved traits and other factors that change as time goes on, like lagged outcomes. It seems like the core problem they're trying to solve is figuring out how to measure that average treatment effect on the treated under these tough selection rules.
Iris: I think it’s important because most of the standard causal methods rely on simpler assumptions about how untreated outcomes behave, like assuming a linear factor model. But this paper uses a nonparametric model that lets the outcomes depend flexibly on their entire history and those time-invariant latent factors, which removes that restriction >
Ruth: Right. That flexibility is key because it lets them deal with selection mechanisms that are way more complex than what models like difference in differences can handle, especially when lagged outcomes affect both who gets treated and what happens next >
Warren: And the authors develop an identification strategy based on these time-invariant latent factors and pretreatment histories, which they claim are informative about those factors even in dynamic settings under weak dependence >
Iris: That’s a big assumption they have to make, that the pretreatment outcome histories actually tell you enough about those latent factors for the identification to work >
Ruth: Yeah, so what they propose is this kernel-based doubly robust estimator for the dynamic average treatment effect on the treated. It uses these flexible models to estimate both what happens to those untreated people and how much of that difference is due to their selection >
Warren: And they’ve combined double cross-fitting with undersmoothing, which they show under suitable regularity conditions it leads to estimators that are root n consistent, asymptotically normal, and asymptotically unbiased >
Iris: That's a strong theoretical claim. They build out the identification strategy through theorem one and then prove the convergence properties in theorem two >
Ruth: It’s a lot of technical stuff there. The simulation study they run shows that this method provides accurate inference across a wide range of data-generating processes, which is pretty encouraging >
Warren: And they apply it to something real, using an example like the U.S. family planning program studied by Bailey in two thousand twelve to reestimate the effect on fertility rates > <ref:2609.17170#pg1>
Iris: It’s interesting because when they looked at that fertility data, they found that accounting for selection based on lagged outcomes is actually crucial for estimating those longer-run treatment effects >
Ruth: So what this means for us is that when we look at things like policy or intervention effects, especially in longitudinal data where people get treated over time, we have to be careful about how selection processes are modeled >
Warren: It suggests that if you ignore the dynamic selection based on lagged outcomes, your estimate of the treatment effect might be biased because you're missing how those past outcomes influence future treatment assignments or results >
Iris: The paper’s limitation is that it relies on those pretreatment outcome histories being informative enough about the latent factors, and they don't explicitly solve the problem if you have a completely different kind of selection mechanism that isn't covered by their framework >
Ruth: Exactly. So while this method handles selection driven by time-invariant factors and covariates, it might not capture every possible way selection can happen in the real world >
Warren: It’s a solid tool for when you have these specific kinds of dynamic confounding, but we gotta remember that the assumptions about those histories being informative are pretty strict >
Iris: And to wrap up this section, the paper gives a clear pipeline: from finding a novel nonparametric way to identify latent heterogeneity, to building that doubly robust estimator with strong convergence properties, and finally validating it on complex fertility data >
Conclusion: Ruth: So we’re wrapping up this look at "Causal Inference under Dynamic Selection and Latent Heterogeneity." Basically, this paper tackles those tough scenarios where you can't just compare groups because when a treatment gets adopted, it depends on things that are unobserved and keep changing over time.
Warren: Right. They developed a way to handle dynamic treatment effects when the selection process is messy, involving both static traits and things that shift as time moves forward.
Iris: The paper focuses on figuring out how to identify the average treatment effect on the treated under those complex rules, using a nonparametric model for untreated outcomes that lets them be flexible with those time-varying factors.
Ruth: And they propose this identification strategy using time-invariant latent factors and pretreatment histories, provided those histories are informative enough about those latent factors.
Warren: The actual estimation tool they build is a kernel-based doubly robust estimator, which uses these flexible models to estimate both the untreated outcomes and how much of the difference comes from selection.
Iris: They show that by combining double cross-fitting with undersmoothing, they get estimators that are root n consistent and asymptotically normal under certain conditions.
Ruth: It’s a big piece of work because it gives us a complete pipeline from identification to estimation with strong mathematical guarantees.
Warren: The application on fertility data shows that ignoring selection based on lagged outcomes can really skew your estimate of how long-term treatment effects are.
Iris: So the main thing here is that for those complex dynamic settings, you need a model that accounts for both the unobserved stuff and the changing covariates when you're trying to measure real causal impacts.
Ruth: Exactly, so this paper shows how to build a more robust tool for understanding things like policy effects or medical interventions when people aren't treated all at once.
Warren: It suggests that if you ignore the dynamic selection based on past outcomes, your estimate of the treatment effect might be biased because you're missing how those past results influence future decisions.
Iris: The paper is quite rigorous in its identification strategy, but it does rely on those pretreatment histories being informative about the latent factors for everything to work as intended.
Ruth: It’s a solid framework, but we still have to keep in mind that the assumptions about those histories being informative are pretty strict when you apply it.
Warren: So this paper is a useful tool for specific kinds of dynamic confounding, but you gotta remember that it might not capture every possible way selection can happen in the real world.
Iris: The authors flag those limitations pretty clearly, so we know where this method stops working and what assumptions need to hold true for the results to be reliable.
Ruth: Exactly. So while this method handles selection driven by time-invariant factors and covariates, it might not cover every possible selection mechanism out there.
Warren: It’s a solid tool for when you have these specific kinds of dynamic confounding, but you gotta remember that it might not capture every possible way selection can happen in the real world.
Iris: The paper is quite rigorous in its identification strategy, but it does rely on those pretreatment histories being informative about the latent factors for everything to work as intended.
Ruth: It’s a big piece of work because it gives us a complete pipeline from identification to estimation with strong mathematical guarantees.
Warren: The application on fertility data shows that ignoring selection based on lagged outcomes can really skew your estimate of how long-term treatment effects are.
Iris: So the main thing here is that for those complex dynamic settings, you need a model that accounts for both the unobserved stuff and the changing covariates when you're trying to measure real causal impacts.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19